Tiiny Pocket's tests produced two clear numbers: about 300 grams, 80GB of memory, and roughly 64 tokens per second when running a 35B mixture-of-experts model. Local AI is no longer confined to desktop experiments, but the 120B model's roughly 16 tokens per second also reminds us that portability does not mean every task will be fast.
What this is
Tiiny Pocket is a palm-sized device for running local AI. Local AI means the model runs on the user's own device rather than uploading data to a cloud server. It uses 32GB of SoC memory and 48GB of dedicated NPU memory; the latter is a computing unit designed specifically for AI workloads.
The tests covered Qwen3.5-9B, Qwen3.6-35B-A3B, and gpt-oss-120b. With an input of roughly 4,000 tokens, the 9B model generated about 75 tokens per second, the 35B model about 64 tokens per second, and the 120B model about 16 tokens per second. The 35B model also completed a four-file Python management tool and passed 10 automated tests, including inventory consistency, order idempotency, and input validation.
Industry view
The important shift is that mixture-of-experts architecture—which activates only part of a model's parameters on each pass—allows a 35B model to maintain relatively high speed on a small device. This creates another viable option for offline coding, document processing, and privacy-sensitive use cases, while lowering the hardware barrier for individual developers looking to experience large models.
The risks are just as clear. The 120B model took more than 30 seconds to complete the same task, revealing a significant trade-off between device memory and speed. This was also only a single-user test, without validation across different thermal conditions, long-term stability, or broader software compatibility. Price, after-sales support, and ecosystem maturity will determine whether Tiiny Pocket becomes a specialized developer tool or a mass-market consumer product.
Impact on regular people
For enterprise IT: Keeping sensitive material on the device has compliance value, but before purchasing, teams should verify that the models required for their workloads can deliver acceptable response times and assess deployment and maintenance costs.
For individual professionals: The portable device can handle light writing, code drafting, and document organization, but complex analysis still requires the cloud or a workstation. Being able to run a model should not be mistaken for being able to replace professional tools.
For the consumer market: Products like this could form a new category between computers and smart speakers. What determines mass adoption will not be the specification sheet, but pricing, battery life, and a software experience that works immediately out of the box.